Static Image Based Pedestrian Detection Methods

Pedestrian detection is a very important research field in the computer vision,which plays a crucial role in intelligent video surveillance,car-assisted driving and pedestrian movement analysis.However,pedestrian detection is one of the hottest research field in artificial intelligence,due to various pedestrian postures and size,changes in terms of light intensity and shooting angles,different cluttered backgrounds and pedestrian occlusions.Besides,the accuracy and real-time requirement are also need to be taken into consideration in pedestrian detection.This paper studies several classic pedestrian detection algorithms,inspired by these classical algorithms,this paper makes improvements to some problems that exist in these algorithms,and proposes three new methods of pedestrian detection,which are verified the validity by experiments.The main contributions of this paper are as follows:(1)This paper makes a brief introduction and summary of the current situation and classic algorithms in pedestrian detection.Moreover,this paper discusses the current problems and difficulties.Meanwhile,the specific process of pedestrian detection and some key technologies involved are introduced in detail.(2)A pedestrian detection method based on Subset-Haar-like features is proposed.By using seven Haar-like feature templates to convolve ten channels of ACF,the target intermediate layer features are generated.Our detector is modelled through the target features in combination with the soft-cascade boosted decision trees.The results experimented on INTIA and Caltech dataset indicate the effectiveness of this proposed method.(3)A pedestrian detection method based on weighted Subset-Haar-like features learnt by LFDA is proposed.The target features consist of the weighted sum of original Subset-Haar-like features,which is learnt by LFDA.Experiment shows that this proposed method reduces the miss rate significantly,while maintaining the detection accuracy.(4)A pedestrian detection method fused by weighted Subset-Haar-like detector and ResNet is proposed.Using weighted Subset-Haar-like model to detect positive training samples of Inria(Caltech)dataset,positive and negative samples used for ResNet training are achieved by Triming bounding boxes,whose IoU Greater than positive threshold and less than negative threshold respectively,the positive bounding boxes are filled with padding 10% operation.Furthermore,ground truth bounding boxes from Inria(Caltech)training set are also trimmed as positive samples for ResNet training.All these images are sent into pre-trained ResNet152 network model fine-tuning,and generate the final model,which used to eliminate false non-pedestrian bounding boxes detected by Ours2(LFDA)model.The effectiveness of this method is indicated by experiment.